Data-driven turbulence modeling for fluid flow and heat transfer in peripheral subchannels of a rod bundle.

This article focuses on improving Reynolds-averaged Navier–Stokes (RANS) turbulence models for predicting mean velocity and scalar fields in turbulent channel flows with complex geometries, specifically peripheral subchannels of rod bundles. It compares multi-dimensional gene expression programming (MGEP), tensor basis neural network (TBNN), and introduces a novel universally interpretable machine learning architecture for turbulent scalar flux (UIML-s). Using high-fidelity direct numerical simulation (DNS) data, the MGEP approach enhances the Reynolds-stress anisotropy tensor modeling, successfully capturing secondary flow structures absent in baseline linear eddy viscosity models, with better performance than TBNN in complex geometries. The UIML-s model improves turbulent scalar flux predictions across a range of Prandtl numbers, reducing normalized root mean squared error from 13.5% to 7.6% compared to conventional gradient diffusion hypotheses. These findings demonstrate the potential of data-driven machine learning calibrations to advance RANS simulations for fluid flow and heat/mass transfer in engineering applications involving complex geometries and varying thermal properties.